SIGN IN SIGN UP

fix(tests): the capability conformance suite encoded this machine's assumptions

Both new tests failed in CI on every platform while passing here, which is
the same mistake the suite was written to catch, made inside the suite.

`readings_claimed_usable_here_actually_resolve` required an observation for
every domain declared usable. A capability describes what this *build* can
read, not what this *machine* has: a CI runner with no GPU produces no gpu
observations and is right to. The workstation this was written on has three
adapters, so it passed.

It now requires simon to say something either way — readings, or the declared
`<domain>.<none>` diagnostic carrying the reason. That is a stronger check
than the original, not a weaker one: silence is the failure, and silence is
the answer this crate exists to prevent.

`every_surface_has_at_least_one_capability` required every surface to be
populated. Settings are derived from registered write handlers and macOS
registers none, so an empty Setting surface there is the truth. It now
asserts the surface and the handler list agree in both directions, which
catches the two failures that matter: capabilities promising writes the
binary cannot perform, and handlers hiding writes it can.

Verified the suite still bites: declaring memory unimplemented while it
resolves fails with the five memory ids that contradict it.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A
Adam Erickson committed
dff34898a32f3a0e8586ecbeb31128e0f4435e96
Parent: 16e4e91